fix(serve): don't classify Gemma 4 closed prefilled thinking block as reasoning

[Anai-Guo]

What

Fixes #46561. With transformers serve and a Gemma 4 model, a normal (thinking-disabled) chat completion comes back with empty content — the whole answer is misclassified as reasoning_content.

Root cause

Gemma 4's chat template prefills an empty, already-closed thinking block at the end of the prompt when thinking is disabled, so the prompt tail is <|channel>thought\n<channel|> → [..., 100, 45518, 107, 101] where start_ids = [100, 45518, 107] and the trailing token 101 is the thinking end token (<channel|>).

_starts_in_thinking() matches start_ids at the tail while tolerating one trailing token (so genuine prefilled openers like DeepSeek-R1 / QwQ that emit <think>\n still match). The closing tag was accepted as that tolerated trailing token, so the heuristic wrongly reported start_in_thinking=True. Because the output contains no thinking markers, parse_reasoning() then takes the "prefilled opener truncated before close" branch and returns content="" with everything in reasoning.

Fix

Thread the already-computed end_id into _starts_in_thinking() and reject the one-trailing-token match when that trailing token is the thinking end token — a closed prefilled block means the prompt is not left inside a thinking block. The trailing == 0 (ends exactly on opener) and trailing == 1 with a non-end token (DeepSeek-R1 / QwQ \n) cases are unchanged.

Tests

Added TestStartsInThinking (CPU-only, no model download) covering: ends-exactly-on-opener, opener + whitespace trailing, the Gemma 4 closed-block case (returns False), the closed block without end_id (back-compat still matches), no-thinking tail, and batched input_ids.

🤖 Generated with Claude Code

[github-actions]

CI Dashboard: View test results in Grafana

[Rocketknight1]

Please stop the random drive-by Claude PRs! We can run Claude Code ourselves if we need it!